SwiftSLU: An Framework for Cost-Efficient Dataset Construction and Boundary-Modifier-Aware Joint Model in Domain-Specific SLU

SLU datasets are highly domain-specific and require a token-level fine-grained annotation, which lead to extremely high cost for manual BIO labeling. In this paper, We introduce a cost-efficient dataset construction approach, demonstrated through the development of VehiCom, a vehicle commands SLU dataset in Chinese built from scratch. We also propose JointVehiCom, a boundary-aware and modifier-aware joint SLU model that excels in accurately identifying entity boundaries in utterances and recognizing entities within auxiliary components, achieving state-of-the-art performance.